A 278-514M Event/s ADC-Less Stochastic Compute-In-Memory Convolution Accelerator for Event Camera

We present a Compute-In-Memory (CIM) convolution accelerator for object tracking applications using event camera, which is a new imaging technology that significantly improves latency and dynamic range over conventional cameras. Previous works proposed an efficient ADC-Iess Stochastic CIM for deep learning but need to store $2^{\mathrm{N}}$ stochastic bits for an n-bit number. We propose to store binary numbers in memory and convert them to stochastic bits by in-situ Stochastic Number Generators on the fly, which reduce the storage requirement by >10x. The CIM macro embeds 32 tiny MAC units per weight and uses an early termination technique to skip unnecessary computation of zeros. The accelerator achieves energy efficiency of 485 TOPS/Wand throughput of 278–514 Mevent/s. The proposed SCIM macro can also be used to accelerate convolution in most deep learning applications.